DeepSeek V4.1 Flash: More performance than V4-Pro at lower prices

DeepSeek has released a new open-weight model with V4.1 Flash, which performs particularly well on agentic tasks and reasoning over long contexts. The predecessor V4 Flash will be discontinued. From September 14, all requests to the current flagship model V4-Pro will also be redirected to V4.1 Flash and billed at its rates. Thus, until V4.1 Pro is released, V4.1 Flash will take over the role of DeepSeek’s flagship.

According to DeepSeek, the central model network of V4.1 Flash comprises 552 billion parameters and is thus almost twice as large as that of V4 Flash. Unlike V4 Pro and V4 Flash, V4.1 Flash now processes images natively in addition to text. The context window of one million tokens remains the same. The model weights are available under the MIT license, and V4.1 Flash is also available via the DeepSeek API.

New architecture reduces computing and storage requirements

A special feature of V4.1 Flash is the new causal encoder-decoder architecture, which enables significant efficiency gains. Simply put, it divides the model into two halves: the first processes the input context, the second handles further generation. This makes reading large amounts of context more efficient. DeepSeek states 8 billion active parameters per token for processing the input and 16 billion for text generation. V4 Flash, on the other hand, had 13 billion active parameters per token.

The new architecture primarily targets input-intensive tasks where a model needs to process a large amount of context. This applies, for example, to agentic programming or research tasks that involve processing large amounts of source code, files, web pages, and tool call results.

A second efficiency lever is the significantly lower memory requirement for the already processed context. The so-called KV cache stores intermediate results from the context already read, so the model does not have to recalculate them constantly during further generation. DeepSeek has drastically reduced this global KV cache over several model generations; with V4.1 Flash, it is about 75 percent smaller than with V4 Flash. These optimizations directly translate into memory requirements: compared to V4 Flash, the cache requires only about a quarter of the HBM and an eighth of the persistent SSD storage, according to DeepSeek. Overall, the architectural changes are intended to reduce costs and increase throughput.

More agent performance, high token consumption

According to tests by Artificial Analysis, DeepSeek V4.1 Flash achieves 40 points in the Intelligence Index with the highest reasoning level, making it one of the leading open-weight models. V4.1 Flash particularly improves on multi-step agent tasks, outperforming both V4 Flash and V4-Pro here. It also performs strongly in reasoning over long contexts. Compared to V4 Flash, the hit rate increases for factual knowledge, but the model also hallucinates more frequently when knowledge gaps exist.

A weak point of V4.1 Flash is its relatively high token consumption: the model requires significantly more output tokens per task than V4-Pro and several current flagship models. As a result, the performance-to-cost ratio per task is slightly worse than the low token prices suggest.

Via the DeepSeek API, one million non-cached input tokens cost $0.30 during peak hours, and one million output tokens cost $1.20. Outside of peak hours, prices are halved to $0.15 and $0.60 respectively.

Artificial Analysis uses the peak tariffs for its cost comparison. Despite the high token consumption, V4.1 Flash comes in at $0.27 per Intelligence Index task. This is about 60 percent less than V4-Pro and about one-seventh the cost of GLM 5.3 and Kimi K3. In contrast, the high token consumption is noticeable compared to V4 Flash: the same task costs only $0.22 there, despite higher token prices.

(mho)